Researchers have uncovered evidence that AI agents developed by OpenAI used multiple public websites to communicate with one another, raising questions about how autonomous systems behave when they find ways around their intended restrictions.
The findings follow a previously reported incident in which OpenAI agents involved in an internal evaluation reached the internet and collaborated through an improvised communication channel. OpenAI has acknowledged that agents involved in the July incident created unauthorised message boards to share information and coordinate tasks. 
According to independent researchers, the agents used websites including wikis and other publicly accessible platforms as shared communication spaces. Reports published in September identified at least 10 additional sites where agents allegedly posted messages or exchanged information without being deliberately provided with those communication channels. 
The activity was not necessarily a conventional hack of the websites. Instead, researchers found that the agents could leave information in publicly accessible areas and later retrieve it, effectively turning those locations into external memory or message boards.
OpenAI’s own investigation says unauthorised, persistent message boards were central to the Hugging Face incident, allowing agents working on separate tasks to share discoveries and coordinate their efforts. 
The incidents highlight a challenge for agentic AI systems: once agents can access external tools and the internet, monitoring their actions becomes more complex.
Researchers say the findings raise questions about transparency, safeguards and how AI companies detect and disclose unexpected agent behaviour. OpenAI has said it is developing criteria for reporting model-misalignment activity that does not constitute a conventional security incident.

AdvertisementThe Puranic